Clustered Federated Learning for Recommendation Systems: Tackling Data Heterogeneity

Recommendation systems have shifted to the federated paradigm (FedRS) to address privacy risks and data silos, yet non-IID data heterogeneity is still a bottleneck limiting FedRS performance. And Clustered Federated Learning (CFL) emerges as a potential solution for recommendation systems. This review attempts to sort out two core CFL frameworks—Iterative Federated Clustering Algorithm (IFCA, alternating cluster identity estimation and model optimization) and SnapCFL (decoupling pre-clustering and dynamic client selection)—and three CFL-based FedRS approaches: user clustering (PerFedRec, FedGWC, FedPCL), item clustering (CoFedRec, ClusterGCF), and user-item co-clustering (CdFed, CPF-GCN), while analyzing challenges like dynamic user preferences, data sparsity/cold-start, and large-system scalability. With the rapid growth of personalized services and increasing privacy concerns, understanding how CFL can enhance model personalization and efficiency has become crucial for advancing privacy-preserving recommender systems. This review tries to present key progress in this field, hoping to provide some reference for readers to grasp the current research status and clarify subsequent directions.